refactor(perf): Remove extra page indirection in Table

[Veykril]

By manually erasing the concrete slot types and dealing with the vtable ourselves

[netlify]

✅ Deploy Preview for salsa-rs canceled.

Name	Link
🔨 Latest commit	4b6dde4
🔍 Latest deploy log	https://app.netlify.com/sites/salsa-rs/deploys/67c9bbd4efbdd700089fc94f

[codspeed-hq]

CodSpeed Performance Report

Merging #710 will improve performances by 4.31%

_{Comparing Veykril:veykril/push-vzrqlltwrwmu (4b6dde4) with master (9d2a978)}

Summary

⚡ 6 improvements
✅ 5 untouched benchmarks

Benchmarks breakdown

	Benchmark	BASE	HEAD	Change
⚡	amortized[SupertypeInput]	4.5 µs	4.3 µs	+4.31%
⚡	new[Input]	11.6 µs	11.2 µs	+4.1%
⚡	new[SupertypeInput]	17.2 µs	16.6 µs	+3.81%
⚡	mutating[10]	14.5 µs	14 µs	+3.55%
⚡	mutating[20]	14.7 µs	14.2 µs	+3.49%
⚡	mutating[30]	14.8 µs	14.3 µs	+3.46%

[Veykril]

Results look promising

[Veykril]

Alright all cleaned up, this will collide with #649

[Veykril]

We could likely give Accumulated a similar treatment to remove the Box<dyn Accumulated>, though that might be a bit trickier since we need to take care of the layout instability of Vec used within somehow

[Veykril]

Given we only ever modify allocated in this function while the allocation_lock is held, could we use Relaxed ordering for the load here ? I believe so as the lock call should already incur a happens-before relationship for the participating threads right?

[Veykril]

In fact it should be fine to make all loads (and the store in the critical section) relaxed due to the sole store happening within the critical section right? (not gonna include this in the PR either way)

[Veykril]

I like the new lower limit, it makes my PRs pop out more now :^)